| Home Resume Academic Schedule Master Thesis |
GPUs changing geometry GPUs improving quality and effects on the pixel level GPUs doing scientific calculation BrookGPU Important Webpages
- GPU's are programmable processors - They have two types of programs (per-vertex and per-pixel) - They use a high-level programming language I'm not intending to tell the whole history of how the GPUs developed, but I want to point out some of their main features. The Cg Tutorial Book (Randina Fernando and Mark J. Kilgard, April 2003) divides the computer graphics hardware in four generation. It only make sense to talk about GPUs programable from the sencond generation (1999-2000), which includes NVIDIA's GeForce 256 and GeForce2, ATI 7500 ... In this generation both, OpenGL and DirectX 7 support hardware vertex transformation. The third generation of GPUs (2001) includes NVIDIA's GeForce3, GeForce4 Ti and ATI's Radeon 8500. Dispite the pixel-level not being powerful enough, this generation has more powerful vertex programing and more available pixel-level configurability. With the fourth generation, (2002 and on) which includes NVIDIA's GeForce Fx family and ATI's Radeon 9700, the programmability opens up in vertex and pixel levels. This is the generation of GPUs where people start thinking about doing not only complicate shading, but scientific computation as well.
Since the second generation, it became possible to send parameters for the vertex shading processor at the GPU in order to apply a function for each vertex. The following figure shows the new feature. In the first sequence (right-top) the CPU is evaluating the function f(x) for each element of vector x[n]. The second sequence (right-bottom) the function has been developed at GPU vertex processor. Only doing this change, the majority of examples presented so far speed up the rendering at least ten times. The figure in the right is a wave simulator applycation developed on Open Inventor, where the GPU is doing a wave displacement function (more ... for details and code).
Although the user had adquired a reasonable control over the vertex processor (since the second generation), it was not until the fourth that he was able to reach better control over the fragment shading process. This generation provides: - support for long shading programs. - high-precision of color and pixel operations (32 bits). - High level shading languages get really interesting. Examples:
With the new GPUs features mentioned above, people started to look at the GPU as a powerful vector coprocessor to the CPU. The intermediate values during a computaion (unsing float buffers) are no longer clamped. Additionally, another good reason to use them as a coprocessor is its parallel nature at the razterization stage (pixel-level). In this way texture-images become matrices of values to do computation. Nowadays, GPUs are being used for linear Algebra computation, signal-images processing, physical simulation and so on ... Limitations - Limited instructions and register space - No branches, but you can use conditionals or multipass - No good enough (so far) to send values back to the CPU Examples:
more examples and appications ...
Now is time to get better idea how people are using GPUs for scientific computation, which is the main goal of this webpage. In order to carry out this ojetive we select Brook for GPU library that was developed at Stanford University. As you have seen, there are already so many application and papers that are using GPUs for computing, but it is difficult to find source code and details, how this whole interaction betwen CPU->GPU and GPU->CPU. This is one of the reason we select BrookGPU. Although this library works on OpenGL and DirectX backend, I will explain briefly the OpenGL part. Introduction questions: What kind of buffer can we use on the GPU for computation? They are additional non-visible rendering buffers called float buffers or pixels buffers (pbuffer). Do I need expecial configuration or support to use them? In the OpenGL case, you will need check if your hardware support certain OpenGL Extensions (see registry). The glxinfo application, on Linux machines, can tell you what extentions are supported in your hardware. Usually you can find it at /usr/X11R6/bin/. BrookGPU requires NVIDIA video card so you have to looking for GL_NV_float_buffer extention. You can check another extentions which not required NVIDIA cards: WGL_ARB_pbuffer GLX_SGIX_pbuffer Note: Before see the next presentation you shoud read BrookGPU webpage. Also you can see GH03-Brook.ppt How BrookGPU works?In this presentation (ppt file, pdf file) I try to explain:- Stream creation - Kernels creation and rendering - Reduction operation Code examples:- Mark J. Harris (RenderTexture class)- NVSDK 6.0 (pbuffer.h pBuffer.cpp) - Patrick Crawley (Unsteady Flow Advection Convolution) Notes:- check this out Request For Comment: EXT_render_target proposal. Thanks Patrick Crawley
Brook for GPU OpenGL HLSL Cg and HLSL FAQ Date: 01/14/2004 Last Update: 04/09/2004 |
|||||||||||||||||||||||||||||
| |||||||||||||||||||||||||||||
|
